With the continuous development of airborne light detection and ranging(LiDAR) technology, the ground point cloud data obtained is trending towards higher density, providing data support for the high-precision construction of terrain models. However, high-density point cloud data often contains a large number of redundant points that do not contribute to terrain expressions. Before constructing the terrain model, it is necessary to perform reasonable data thinning on the high-density point cloud data. The existing thinning algorithms face low efficiency and long computation time in thinning massive high-density point cloud data, which cannot meet the practical needs of time-critical application fields such as emergency rescue. In order to promote the deepening application of airborne LiDAR technology in time-critical fields such as emergency rescue, a fast thinning method for airborne LiDAR point clouds based on triangulated irregular network thinning algorithm is proposed for terrain model construction, aiming to optimize the computational efficiency of existing thinning algorithms.

In this method, the original ground point cloud was initially thinned based on a multi-scale grid, and the differences in grid scales were used to quickly extract terrain feature points and boundary points. These points were then used as seed points for constructing an initial triangulated network that fit the macroscopic terrain features, solving the problem of excessive redundant points and the inability to effectively control the network size in existing methods. Secondly, during the densification process of the triangulated network, the centroid point of the newly constructed triangle in the triangulated network was iteratively selected as the insertion point for evaluation, and the insertion points that were farther away from the existing triangulated network nodes were uniformly selected. At the same time, the grid method was used to avoid point by point retrieval operations within the triangle of the triangulated network, greatly improving the efficiency of triangulated network densification. Finally, the densified triangulated network was subjected to redundant point filtering, and the filtered triangulation nodes were used as the thinning result.

In order to verify the influence of various parameters on the thinning results and computational efficiency in this method, comparative experiments were conducted on three parameters: g_\textmax (large grid size) , g_\min (small grid size), and D (height difference threshold). The experimental results proved that g_\textmax were used to control the construction speed and accuracy of the initial triangulated network in the algorithm, g_\min controlled the number of points involved in elevation deviation detection during the triangulation densification process, and parameter D controlled the number of nodes inserted into the triangulated network through elevation deviation detection. Secondly, to verify the effectiveness of this method, experimental comparisons were made with existing optimization methods in terms of computation time, thinning rate, and root mean square error of model elevation. As shown in Fig.9, this method optimized the computation time to between 1.5% and 11.0% of the original time compared to existing optimization methods, while ensuring the accuracy of constructing terrain models from point clouds after thinning. This greatly improved the computational efficiency of the thinning algorithm.

Therefore, the research results have achieved significant optimization of operational efficiency compared to existing point cloud thinning algorithms, while ensuring the accuracy of thinning. It has strong engineering application value and can promote the deepening application of airborne LiDAR technology in time-critical fields such as emergency rescue. In addition, as the proposed method uses a single triangle in a triangulated network as an independent unit for recursion, it can be utilized for parallel computing design in future work, further improving the efficiency of thinning computation.